CT metal artifact reduction using MR image patches

Abstract

Metal implants give rise to metal artifacts in computed tomography (CT) images, which may lead to diagnostic errors and erroneous CT number estimates when the CT is used for radiation therapy planning. Methods for reducing metal artifacts by exploiting the anatomical information provided by coregistered magnetic resonance (MR) images are of great potential value, but remain technically challenging due to the poor contrast between bone and air on the MR image. In this paper, we present a novel MR-based algorithm for automatic CT metal artifact reduction (MAR), referred to as kerMAR. It combines kernel regression on known CT value/MR patch pairs in the uncorrupted patient volume with a forward model of the artifact corrupted values to estimate CT replacement values. In contrast to pseudo-CT generation that builds on multi-patient modelling, the algorithm requires no MR intensity normalisation or atlas registration. Image results for 7 head-and-neck radiation therapy patients with T1-weighted images acquired in the same fixation as the RT planning CT suggest a potential for more complete MAR close to the metal implants than the oMAR algorithm (Philips) used clinically. Our results further show improved performance in air and bone regions as compared to other MR-based MAR algorithms. In addition, we experimented with using kerMAR to define a prior for iterative reconstruction with the maximum likelihood transmission reconstruction algorithm, however with no apparent improvements

MR‐based CT metal artifact reduction for head‐and‐neck photon, electron, and proton radiotherapy

“…kerMAR, however, has certain potential drawbacks that one should bear in mind . In particular, when the MRI and CT are not well aligned, the kerMAR may introduce artifacts as the MR‐based prediction is compromised.…”

“…along with further details. kerMAR is an image space, Bayesian inference algorithm that uses kernel regression on aligned uncorrupted CT values and cuboidal MRI patches (vectors of MRI voxel intensities from local spatial contexts) in the patient volume. It estimates a prior distribution of the true CT value y given the corresponding MRI patch m , p ( y | m ).…”

“…Assuming additive Gaussian artifact noise and given an observed m centered on a corrupted location as well as the corresponding corrupted CT value t , the posterior distribution p ( y | t, m ) is then constructed. Calculating the expectation value of y over this distribution yields the final CT value estimate . Not only relying on CT information, a potential for artifact reduction in highly corrupted regions is possible. oMAR is an iterative algorithm that combines image processing and projection replacement.…”

“… Manual override techniques address the metal artifacts by visually replacing artifact‐corrupted regions with a bulk CT value. Since the oral cavity is largely water equivalent, a plausible CT value is 0 Hounsfield Units (HU), leading to a water override. How this override is performed depends on the local clinical practice; in our approach, we replaced obviously corrupted soft tissue regions as well as severely corrupted high intensity regions by this reproducible, generic value.…”

“…We recently presented a novel example of one such alternative, which, in addition to the corrupted CT information, incorporates complementary image information from magnetic resonance imaging (MRI) using kernel regression (ker-MAR). 18 Taking advantage of the less artifact-corrupted MRI that can be acquired for H&N RT to aid in tumor and OAR delineation, kerMAR uses kernel regression on CT value/MR image patch pairs along with a noise model of the CT artifacts to estimate the true CT values underlying the artifacts. This leads to potential improvements, in particular close to the metal implants where purely CT-based algorithms are the least effective.…”